llvmpipe: Temporary workaround to prevent segfault on array textures.
[mesa.git] / src / gallium / drivers / llvmpipe / lp_setup.c
index 2597fa8f71b09f8081bf7c63a353022fa729b4d0..8c9f0c646539b113b3b76e300bbd9bd23bec6602 100644 (file)
 #include "util/u_inlines.h"
 #include "util/u_memory.h"
 #include "util/u_pack_color.h"
+#include "draw/draw_pipe.h"
 #include "lp_context.h"
 #include "lp_memory.h"
 #include "lp_scene.h"
-#include "lp_scene_queue.h"
 #include "lp_texture.h"
 #include "lp_debug.h"
 #include "lp_fence.h"
 #include "draw/draw_vbuf.h"
 
 
-static void set_scene_state( struct lp_setup_context *, enum setup_state );
+static boolean set_scene_state( struct lp_setup_context *, enum setup_state,
+                             const char *reason);
+static boolean try_update_scene_state( struct lp_setup_context *setup );
 
 
-struct lp_scene *
-lp_setup_get_current_scene(struct lp_setup_context *setup)
+static void
+lp_setup_get_empty_scene(struct lp_setup_context *setup)
 {
-   if (!setup->scene) {
-      set_scene_state( setup, SETUP_EMPTY );
-   }
-   return setup->scene;
-}
+   struct llvmpipe_context *lp = llvmpipe_context(setup->pipe);
+   boolean discard = lp->rasterizer ? lp->rasterizer->rasterizer_discard : FALSE;
 
+   assert(setup->scene == NULL);
 
-/**
- * Check if the size of the current scene has exceeded the limit.
- * If so, flush/render it.
- */
-static void
-setup_check_scene_size_and_flush(struct lp_setup_context *setup)
-{
-   if (setup->scene) {
-      struct lp_scene *scene = lp_setup_get_current_scene(setup);
-      unsigned size = lp_scene_get_size(scene);
+   setup->scene_idx++;
+   setup->scene_idx %= Elements(setup->scenes);
 
-      if (size > LP_MAX_SCENE_SIZE) {
-         /*printf("LLVMPIPE: scene size = %u, flushing.\n", size);*/
-         set_scene_state( setup, SETUP_FLUSHED );
-         /*assert(lp_scene_get_size(scene) == 0);*/
-      }
+   setup->scene = setup->scenes[setup->scene_idx];
+
+   if (setup->scene->fence) {
+      if (LP_DEBUG & DEBUG_SETUP)
+         debug_printf("%s: wait for scene %d\n",
+                      __FUNCTION__, setup->scene->fence->id);
+
+      lp_fence_wait(setup->scene->fence);
    }
+
+   lp_scene_begin_binning(setup->scene, &setup->fb, discard);
+   
 }
 
 
@@ -96,7 +94,7 @@ first_triangle( struct lp_setup_context *setup,
                 const float (*v1)[4],
                 const float (*v2)[4])
 {
-   set_scene_state( setup, SETUP_ACTIVE );
+   assert(setup->state == SETUP_ACTIVE);
    lp_setup_choose_triangle( setup );
    setup->triangle( setup, v0, v1, v2 );
 }
@@ -106,7 +104,7 @@ first_line( struct lp_setup_context *setup,
            const float (*v0)[4],
            const float (*v1)[4])
 {
-   set_scene_state( setup, SETUP_ACTIVE );
+   assert(setup->state == SETUP_ACTIVE);
    lp_setup_choose_line( setup );
    setup->line( setup, v0, v1 );
 }
@@ -115,18 +113,22 @@ static void
 first_point( struct lp_setup_context *setup,
             const float (*v0)[4])
 {
-   set_scene_state( setup, SETUP_ACTIVE );
+   assert(setup->state == SETUP_ACTIVE);
    lp_setup_choose_point( setup );
    setup->point( setup, v0 );
 }
 
-static void reset_context( struct lp_setup_context *setup )
+void lp_setup_reset( struct lp_setup_context *setup )
 {
+   unsigned i;
+
    LP_DBG(DEBUG_SETUP, "%s\n", __FUNCTION__);
 
    /* Reset derived state */
-   setup->constants.stored_size = 0;
-   setup->constants.stored_data = NULL;
+   for (i = 0; i < Elements(setup->constants); ++i) {
+      setup->constants[i].stored_size = 0;
+      setup->constants[i].stored_data = NULL;
+   }
    setup->fs.stored = NULL;
    setup->dirty = ~0;
 
@@ -135,8 +137,7 @@ static void reset_context( struct lp_setup_context *setup )
 
    /* Reset some state:
     */
-   setup->clear.flags = 0;
-   setup->clear.clearzs.clearzs_mask = 0;
+   memset(&setup->clear, 0, sizeof setup->clear);
 
    /* Have an explicit "start-binning" call and get rid of this
     * pointer twiddling?
@@ -151,22 +152,59 @@ static void reset_context( struct lp_setup_context *setup )
 static void
 lp_setup_rasterize_scene( struct lp_setup_context *setup )
 {
-   struct lp_scene *scene = lp_setup_get_current_scene(setup);
+   struct lp_scene *scene = setup->scene;
+   struct llvmpipe_screen *screen = llvmpipe_screen(scene->pipe->screen);
+
+   lp_scene_end_binning(scene);
+
+   lp_fence_reference(&setup->last_fence, scene->fence);
 
-   lp_scene_rasterize(scene, setup->rast);
+   if (setup->last_fence)
+      setup->last_fence->issued = TRUE;
 
-   reset_context( setup );
+   pipe_mutex_lock(screen->rast_mutex);
+   lp_rast_queue_scene(screen->rast, scene);
+   lp_rast_finish(screen->rast);
+   pipe_mutex_unlock(screen->rast_mutex);
+
+   lp_scene_end_rasterization(setup->scene);
+   lp_setup_reset( setup );
 
    LP_DBG(DEBUG_SETUP, "%s done \n", __FUNCTION__);
 }
 
 
 
-static void
+static boolean
 begin_binning( struct lp_setup_context *setup )
 {
-   struct lp_scene *scene = lp_setup_get_current_scene(setup);
+   struct lp_scene *scene = setup->scene;
    boolean need_zsload = FALSE;
+   boolean ok;
+   unsigned i, j;
+
+   assert(scene);
+   assert(scene->fence == NULL);
+
+   /* Always create a fence:
+    */
+   scene->fence = lp_fence_create(MAX2(1, setup->num_threads));
+   if (!scene->fence)
+      return FALSE;
+
+   /* Initialize the bin flags and x/y coords:
+    */
+   for (i = 0; i < scene->tiles_x; i++) {
+      for (j = 0; j < scene->tiles_y; j++) {
+         scene->tile[i][j].x = i;
+         scene->tile[i][j].y = j;
+      }
+   }
+
+   ok = try_update_scene_state(setup);
+   if (!ok)
+      return FALSE;
+
    if (setup->fb.zsbuf &&
        ((setup->clear.flags & PIPE_CLEAR_DEPTHSTENCIL) != PIPE_CLEAR_DEPTHSTENCIL) &&
         util_format_is_depth_and_stencil(setup->fb.zsbuf->format))
@@ -178,10 +216,11 @@ begin_binning( struct lp_setup_context *setup )
 
    if (setup->fb.nr_cbufs) {
       if (setup->clear.flags & PIPE_CLEAR_COLOR) {
-         lp_scene_bin_everywhere( scene, 
-                                 lp_rast_clear_color, 
-                                 setup->clear.color );
-         scene->has_color_clear = TRUE;
+         ok = lp_scene_bin_everywhere( scene, 
+                                       LP_RAST_OP_CLEAR_COLOR, 
+                                       setup->clear.color );
+         if (!ok)
+            return FALSE;
       }
    }
 
@@ -189,13 +228,33 @@ begin_binning( struct lp_setup_context *setup )
       if (setup->clear.flags & PIPE_CLEAR_DEPTHSTENCIL) {
          if (!need_zsload)
             scene->has_depthstencil_clear = TRUE;
-         lp_scene_bin_everywhere( scene,
-                                  lp_rast_clear_zstencil,
-                                  lp_rast_arg_clearzs(&setup->clear.clearzs) );
+
+         ok = lp_scene_bin_everywhere( scene,
+                                       LP_RAST_OP_CLEAR_ZSTENCIL,
+                                       lp_rast_arg_clearzs(
+                                          setup->clear.zsvalue,
+                                          setup->clear.zsmask));
+         if (!ok)
+            return FALSE;
       }
    }
 
+   for (i = 0; i < PIPE_QUERY_TYPES; ++i) {
+      if (setup->active_query[i]) {
+         ok = lp_scene_bin_everywhere( scene,
+                                       LP_RAST_OP_BEGIN_QUERY,
+                                       lp_rast_arg_query(setup->active_query[i]) );
+         if (!ok)
+            return FALSE;
+      }
+   }
+
+   setup->clear.flags = 0;
+   setup->clear.zsmask = 0;
+   setup->clear.zsvalue = 0;
+
    LP_DBG(DEBUG_SETUP, "%s done\n", __FUNCTION__);
+   return TRUE;
 }
 
 
@@ -204,93 +263,96 @@ begin_binning( struct lp_setup_context *setup )
  *
  * TODO: fast path for fullscreen clears and no triangles.
  */
-static void
+static boolean
 execute_clears( struct lp_setup_context *setup )
 {
    LP_DBG(DEBUG_SETUP, "%s\n", __FUNCTION__);
 
-   begin_binning( setup );
-   lp_setup_rasterize_scene( setup );
+   return begin_binning( setup );
 }
 
+const char *states[] = {
+   "FLUSHED",
+   "CLEARED",
+   "ACTIVE "
+};
 
-static void
+
+static boolean
 set_scene_state( struct lp_setup_context *setup,
-                 enum setup_state new_state )
+                 enum setup_state new_state,
+                 const char *reason)
 {
    unsigned old_state = setup->state;
 
    if (old_state == new_state)
-      return;
-       
-   LP_DBG(DEBUG_SETUP, "%s old %d new %d\n", __FUNCTION__, old_state, new_state);
-
-   switch (new_state) {
-   case SETUP_EMPTY:
-      assert(old_state == SETUP_FLUSHED);
-      assert(setup->scene == NULL);
+      return TRUE;
+   
+   if (LP_DEBUG & DEBUG_SCENE) {
+      debug_printf("%s old %s new %s%s%s\n",
+                   __FUNCTION__,
+                   states[old_state],
+                   states[new_state],
+                   (new_state == SETUP_FLUSHED) ? ": " : "",
+                   (new_state == SETUP_FLUSHED) ? reason : "");
+
+      if (new_state == SETUP_FLUSHED && setup->scene)
+         lp_debug_draw_bins_by_cmd_length(setup->scene);
+   }
 
-      /* wait for a free/empty scene
-       */
-      setup->scene = lp_scene_dequeue(setup->empty_scenes, TRUE);
-      assert(lp_scene_is_empty(setup->scene));
-      lp_scene_begin_binning(setup->scene,
-                             &setup->fb );
-      break;
+   /* wait for a free/empty scene
+    */
+   if (old_state == SETUP_FLUSHED) 
+      lp_setup_get_empty_scene(setup);
 
+   switch (new_state) {
    case SETUP_CLEARED:
-      assert(old_state == SETUP_EMPTY);
-      assert(setup->scene != NULL);
       break;
 
    case SETUP_ACTIVE:
-      assert(old_state == SETUP_EMPTY ||
-             old_state == SETUP_CLEARED);
-      assert(setup->scene != NULL);
-      begin_binning( setup );
+      if (!begin_binning( setup ))
+         goto fail;
       break;
 
    case SETUP_FLUSHED:
       if (old_state == SETUP_CLEARED)
-         execute_clears( setup );
-      else
-         lp_setup_rasterize_scene( setup );
+         if (!execute_clears( setup ))
+            goto fail;
+
+      lp_setup_rasterize_scene( setup );
       assert(setup->scene == NULL);
       break;
 
    default:
       assert(0 && "invalid setup state mode");
+      goto fail;
    }
 
    setup->state = new_state;
+   return TRUE;
+
+fail:
+   if (setup->scene) {
+      lp_scene_end_rasterization(setup->scene);
+      setup->scene = NULL;
+   }
+
+   setup->state = SETUP_FLUSHED;
+   lp_setup_reset( setup );
+   return FALSE;
 }
 
 
-/**
- * \param flags  bitmask of PIPE_FLUSH_x flags
- */
 void
 lp_setup_flush( struct lp_setup_context *setup,
-                unsigned flags )
+                struct pipe_fence_handle **fence,
+                const char *reason)
 {
-   LP_DBG(DEBUG_SETUP, "%s\n", __FUNCTION__);
+   set_scene_state( setup, SETUP_FLUSHED, reason );
 
-   if (setup->scene) {
-      struct lp_scene *scene = lp_setup_get_current_scene(setup);
-      union lp_rast_cmd_arg dummy = {0};
-
-      if (flags & (PIPE_FLUSH_SWAPBUFFERS |
-                   PIPE_FLUSH_FRAME)) {
-         /* Store colors in the linear color buffer(s).
-          * If we don't do this here, we'll end up converting the tiled
-          * data to linear in the texture_unmap() function, which will
-          * not be a parallel/threaded operation as here.
-          */
-         lp_scene_bin_everywhere(scene, lp_rast_store_color, dummy);
-      }
+   if (fence) {
+      lp_fence_reference((struct lp_fence **)fence, setup->last_fence);
    }
-
-   set_scene_state( setup, SETUP_FLUSHED );
 }
 
 
@@ -302,7 +364,7 @@ lp_setup_bind_framebuffer( struct lp_setup_context *setup,
 
    /* Flush any old scene.
     */
-   set_scene_state( setup, SETUP_FLUSHED );
+   set_scene_state( setup, SETUP_FLUSHED, __FUNCTION__ );
 
    /*
     * Ensure the old scene is not reused.
@@ -313,81 +375,52 @@ lp_setup_bind_framebuffer( struct lp_setup_context *setup,
     * scene.
     */
    util_copy_framebuffer_state(&setup->fb, fb);
+   setup->framebuffer.x0 = 0;
+   setup->framebuffer.y0 = 0;
+   setup->framebuffer.x1 = fb->width-1;
+   setup->framebuffer.y1 = fb->height-1;
+   setup->dirty |= LP_SETUP_NEW_SCISSOR;
 }
 
 
-void
-lp_setup_clear( struct lp_setup_context *setup,
-                const float *color,
-                double depth,
-                unsigned stencil,
-                unsigned flags )
+static boolean
+lp_setup_try_clear( struct lp_setup_context *setup,
+                    const union pipe_color_union *color,
+                    double depth,
+                    unsigned stencil,
+                    unsigned flags )
 {
-   struct lp_scene *scene = lp_setup_get_current_scene(setup);
+   uint32_t zsmask = 0;
+   uint32_t zsvalue = 0;
+   union lp_rast_cmd_arg color_arg;
    unsigned i;
-   boolean full_zs_clear = TRUE;
-   uint32_t mask = 0;
 
    LP_DBG(DEBUG_SETUP, "%s state %d\n", __FUNCTION__, setup->state);
 
-
    if (flags & PIPE_CLEAR_COLOR) {
-      for (i = 0; i < 4; ++i)
-         setup->clear.color.clear_color[i] = float_to_ubyte(color[i]);
+      for (i = 0; i < 4; i++)
+         color_arg.clear_color.i[i] = color->i[i];
    }
 
    if (flags & PIPE_CLEAR_DEPTHSTENCIL) {
-      if (setup->fb.zsbuf &&
-          ((flags & PIPE_CLEAR_DEPTHSTENCIL) != PIPE_CLEAR_DEPTHSTENCIL) &&
-           util_format_is_depth_and_stencil(setup->fb.zsbuf->format))
-         full_zs_clear = FALSE;
-
-      if (full_zs_clear) {
-         setup->clear.clearzs.clearzs_value =
-            util_pack_z_stencil(setup->fb.zsbuf->format,
-                                depth,
-                                stencil);
-         setup->clear.clearzs.clearzs_mask = 0xffffffff;
-      }
-      else {
-         /* hmm */
-         uint32_t tmpval;
-         if (flags & PIPE_CLEAR_DEPTH) {
-            tmpval = util_pack_z(setup->fb.zsbuf->format,
-                                 depth);
-            switch (setup->fb.zsbuf->format) {
-            case PIPE_FORMAT_Z24_UNORM_S8_USCALED:
-               mask = 0xffffff;
-               break;
-            case PIPE_FORMAT_S8_USCALED_Z24_UNORM:
-               mask = 0xffffff00;
-               break;
-            default:
-               assert(0);
-            }
-         }
-         else {
-            switch (setup->fb.zsbuf->format) {
-            case PIPE_FORMAT_Z24_UNORM_S8_USCALED:
-               mask = 0xff000000;
-               tmpval = stencil << 24;
-               break;
-            case PIPE_FORMAT_S8_USCALED_Z24_UNORM:
-               mask = 0xff;
-               tmpval = stencil;
-               break;
-            default:
-               assert(0);
-               tmpval = 0;
-            }
-         }
-         setup->clear.clearzs.clearzs_mask |= mask;
-         setup->clear.clearzs.clearzs_value =
-            (setup->clear.clearzs.clearzs_value & ~mask) | (tmpval & mask);
-      }
+      uint32_t zmask = (flags & PIPE_CLEAR_DEPTH) ? ~0 : 0;
+      uint32_t smask = (flags & PIPE_CLEAR_STENCIL) ? ~0 : 0;
+
+      zsvalue = util_pack_z_stencil(setup->fb.zsbuf->format,
+                                    depth,
+                                    stencil);
+
+
+      zsmask = util_pack_mask_z_stencil(setup->fb.zsbuf->format,
+                                        zmask,
+                                        smask);
+
+      zsvalue &= zsmask;
    }
 
    if (setup->state == SETUP_ACTIVE) {
+      struct lp_scene *scene = setup->scene;
+
       /* Add the clear to existing scene.  In the unusual case where
        * both color and depth-stencil are being cleared when there's
        * already been some rendering, we could discard the currently
@@ -395,24 +428,18 @@ lp_setup_clear( struct lp_setup_context *setup,
        * a common usage.
        */
       if (flags & PIPE_CLEAR_COLOR) {
-         lp_scene_bin_everywhere( scene, 
-                                  lp_rast_clear_color,
-                                  setup->clear.color );
-         scene->has_color_clear = TRUE;
+         if (!lp_scene_bin_everywhere( scene, 
+                                       LP_RAST_OP_CLEAR_COLOR,
+                                       color_arg ))
+            return FALSE;
       }
 
       if (flags & PIPE_CLEAR_DEPTHSTENCIL) {
-         if (full_zs_clear)
-            scene->has_depthstencil_clear = TRUE;
-         else
-            setup->clear.clearzs.clearzs_mask = mask;
-         lp_scene_bin_everywhere( scene,
-                                  lp_rast_clear_zstencil,
-                                  lp_rast_arg_clearzs(&setup->clear.clearzs) );
-
-
+         if (!lp_scene_bin_everywhere( scene,
+                                       LP_RAST_OP_CLEAR_ZSTENCIL,
+                                       lp_rast_arg_clearzs(zsvalue, zsmask) ))
+            return FALSE;
       }
-
    }
    else {
       /* Put ourselves into the 'pre-clear' state, specifically to try
@@ -420,39 +447,43 @@ lp_setup_clear( struct lp_setup_context *setup,
        * buffers which the app or state-tracker might issue
        * separately.
        */
-      set_scene_state( setup, SETUP_CLEARED );
+      set_scene_state( setup, SETUP_CLEARED, __FUNCTION__ );
 
       setup->clear.flags |= flags;
+
+      if (flags & PIPE_CLEAR_DEPTHSTENCIL) {
+         setup->clear.zsmask |= zsmask;
+         setup->clear.zsvalue =
+            (setup->clear.zsvalue & ~zsmask) | (zsvalue & zsmask);
+      }
+
+      if (flags & PIPE_CLEAR_COLOR) {
+         memcpy(&setup->clear.color.clear_color,
+                &color_arg,
+                sizeof setup->clear.color.clear_color);
+      }
    }
+   
+   return TRUE;
 }
 
-
-/**
- * Emit a fence.
- */
-struct pipe_fence_handle *
-lp_setup_fence( struct lp_setup_context *setup )
+void
+lp_setup_clear( struct lp_setup_context *setup,
+                const union pipe_color_union *color,
+                double depth,
+                unsigned stencil,
+                unsigned flags )
 {
-   if (setup->num_threads == 0) {
-      return NULL;
-   }
-   else {
-      struct lp_scene *scene = lp_setup_get_current_scene(setup);
-      const unsigned rank = lp_scene_get_num_bins( scene ); /* xxx */
-      struct lp_fence *fence = lp_fence_create(rank);
+   if (!lp_setup_try_clear( setup, color, depth, stencil, flags )) {
+      lp_setup_flush(setup, NULL, __FUNCTION__);
 
-      LP_DBG(DEBUG_SETUP, "%s rank %u\n", __FUNCTION__, rank);
+      if (!lp_setup_try_clear( setup, color, depth, stencil, flags ))
+         assert(0);
+   }
+}
 
-      set_scene_state( setup, SETUP_ACTIVE );
 
-      /* insert the fence into all command bins */
-      lp_scene_bin_everywhere( scene,
-                               lp_rast_fence,
-                               lp_rast_arg_fence(fence) );
 
-      return (struct pipe_fence_handle *) fence;
-   }
-}
 
 
 void 
@@ -460,28 +491,54 @@ lp_setup_set_triangle_state( struct lp_setup_context *setup,
                              unsigned cull_mode,
                              boolean ccw_is_frontface,
                              boolean scissor,
-                             boolean gl_rasterization_rules)
+                             boolean half_pixel_center,
+                             boolean bottom_edge_rule)
 {
    LP_DBG(DEBUG_SETUP, "%s\n", __FUNCTION__);
 
    setup->ccw_is_frontface = ccw_is_frontface;
    setup->cullmode = cull_mode;
    setup->triangle = first_triangle;
-   setup->scissor_test = scissor;
-   setup->pixel_offset = gl_rasterization_rules ? 0.5f : 0.0f;
+   setup->pixel_offset = half_pixel_center ? 0.5f : 0.0f;
+   setup->bottom_edge_rule = bottom_edge_rule;
+
+   if (setup->scissor_test != scissor) {
+      setup->dirty |= LP_SETUP_NEW_SCISSOR;
+      setup->scissor_test = scissor;
+   }
 }
 
+void 
+lp_setup_set_line_state( struct lp_setup_context *setup,
+                        float line_width)
+{
+   LP_DBG(DEBUG_SETUP, "%s\n", __FUNCTION__);
 
+   setup->line_width = line_width;
+}
 
-void
-lp_setup_set_fs_inputs( struct lp_setup_context *setup,
-                        const struct lp_shader_input *input,
-                        unsigned nr )
+void 
+lp_setup_set_point_state( struct lp_setup_context *setup,
+                          float point_size,                          
+                          boolean point_size_per_vertex,
+                          uint sprite_coord_enable,
+                          uint sprite_coord_origin)
 {
-   LP_DBG(DEBUG_SETUP, "%s %p %u\n", __FUNCTION__, (void *) input, nr);
+   LP_DBG(DEBUG_SETUP, "%s\n", __FUNCTION__);
 
-   memcpy( setup->fs.input, input, nr * sizeof input[0] );
-   setup->fs.nr_inputs = nr;
+   setup->point_size = point_size;
+   setup->sprite_coord_enable = sprite_coord_enable;
+   setup->sprite_coord_origin = sprite_coord_origin;
+   setup->point_size_per_vertex = point_size_per_vertex;
+}
+
+void
+lp_setup_set_setup_variant( struct lp_setup_context *setup,
+                           const struct lp_setup_variant *variant)
+{
+   LP_DBG(DEBUG_SETUP, "%s\n", __FUNCTION__);
+   
+   setup->setup.variant = variant;
 }
 
 void
@@ -498,12 +555,21 @@ lp_setup_set_fs_variant( struct lp_setup_context *setup,
 
 void
 lp_setup_set_fs_constants(struct lp_setup_context *setup,
-                          struct pipe_resource *buffer)
+                          unsigned num,
+                          struct pipe_constant_buffer *buffers)
 {
-   LP_DBG(DEBUG_SETUP, "%s %p\n", __FUNCTION__, (void *) buffer);
+   unsigned i;
 
-   pipe_resource_reference(&setup->constants.current, buffer);
+   LP_DBG(DEBUG_SETUP, "%s %p\n", __FUNCTION__, (void *) buffers);
 
+   assert(num <= Elements(setup->constants));
+
+   for (i = 0; i < num; ++i) {
+      util_copy_constant_buffer(&setup->constants[i].current, &buffers[i]);
+   }
+   for (; i < Elements(setup->constants); i++) {
+      util_copy_constant_buffer(&setup->constants[i].current, NULL);
+   }
    setup->dirty |= LP_SETUP_NEW_CONSTANTS;
 }
 
@@ -557,10 +623,11 @@ lp_setup_set_scissor( struct lp_setup_context *setup,
 
    assert(scissor);
 
-   if (memcmp(&setup->scissor.current, scissor, sizeof(*scissor)) != 0) {
-      setup->scissor.current = *scissor; /* struct copy */
-      setup->dirty |= LP_SETUP_NEW_SCISSOR;
-   }
+   setup->scissor.x0 = scissor->minx;
+   setup->scissor.x1 = scissor->maxx-1;
+   setup->scissor.y0 = scissor->miny;
+   setup->scissor.y1 = scissor->maxy-1;
+   setup->dirty |= LP_SETUP_NEW_SCISSOR;
 }
 
 
@@ -571,6 +638,15 @@ lp_setup_set_flatshade_first( struct lp_setup_context *setup,
    setup->flatshade_first = flatshade_first;
 }
 
+void
+lp_setup_set_rasterizer_discard( struct lp_setup_context *setup,
+                                 boolean rasterizer_discard )
+{
+   if (setup->rasterizer_discard != rasterizer_discard) {
+      setup->rasterizer_discard = rasterizer_discard;
+      set_scene_state( setup, SETUP_FLUSHED, __FUNCTION__ );
+   }
+}
 
 void 
 lp_setup_set_vertex_info( struct lp_setup_context *setup,
@@ -594,45 +670,117 @@ lp_setup_set_fragment_sampler_views(struct lp_setup_context *setup,
 
    LP_DBG(DEBUG_SETUP, "%s\n", __FUNCTION__);
 
-   assert(num <= PIPE_MAX_SAMPLERS);
+   assert(num <= PIPE_MAX_SHADER_SAMPLER_VIEWS);
 
-   for (i = 0; i < PIPE_MAX_SAMPLERS; i++) {
+   for (i = 0; i < PIPE_MAX_SHADER_SAMPLER_VIEWS; i++) {
       struct pipe_sampler_view *view = i < num ? views[i] : NULL;
 
-      if(view) {
-         struct pipe_resource *tex = view->texture;
-         struct llvmpipe_resource *lp_tex = llvmpipe_resource(tex);
+      if (view) {
+         struct pipe_resource *res = view->texture;
+         struct llvmpipe_resource *lp_tex = llvmpipe_resource(res);
          struct lp_jit_texture *jit_tex;
          jit_tex = &setup->fs.current.jit_context.textures[i];
-         jit_tex->width = tex->width0;
-         jit_tex->height = tex->height0;
-         jit_tex->depth = tex->depth0;
-         jit_tex->last_level = tex->last_level;
 
          /* We're referencing the texture's internal data, so save a
           * reference to it.
           */
-         pipe_resource_reference(&setup->fs.current_tex[i], tex);
+         pipe_resource_reference(&setup->fs.current_tex[i], res);
 
          if (!lp_tex->dt) {
-            /* regular texture - setup array of mipmap level pointers */
+            /* regular texture - setup array of mipmap level offsets */
+            void *mip_ptr;
             int j;
-            for (j = 0; j <= tex->last_level; j++) {
-               jit_tex->data[j] =
-                  llvmpipe_get_texture_image_all(lp_tex, j, LP_TEX_USAGE_READ,
-                                                 LP_TEX_LAYOUT_LINEAR);
-               jit_tex->row_stride[j] = lp_tex->row_stride[j];
-               jit_tex->img_stride[j] = lp_tex->img_stride[j];
-
-               if (!jit_tex->data[j]) {
-                  /* out of memory - use dummy tile memory */
-                  jit_tex->data[j] = lp_get_dummy_tile();
-                  jit_tex->width = TILE_SIZE;
-                  jit_tex->height = TILE_SIZE;
-                  jit_tex->depth = 1;
-                  jit_tex->last_level = 0;
-                  jit_tex->row_stride[j] = 0;
-                  jit_tex->img_stride[j] = 0;
+            unsigned first_level = 0;
+            unsigned last_level = 0;
+
+            if (llvmpipe_resource_is_texture(res)) {
+               first_level = view->u.tex.first_level;
+               last_level = view->u.tex.last_level;
+               assert(first_level <= last_level);
+               assert(last_level <= res->last_level);
+
+               /*
+                * The complexity here is only necessary for depth textures which
+                * still are tiled.
+                */
+               mip_ptr = llvmpipe_get_texture_image_all(lp_tex, first_level,
+                                                        LP_TEX_USAGE_READ,
+                                                        LP_TEX_LAYOUT_LINEAR);
+               jit_tex->base = lp_tex->linear_img.data;
+            }
+            else {
+               mip_ptr = lp_tex->data;
+               jit_tex->base = mip_ptr;
+            }
+
+            if ((LP_PERF & PERF_TEX_MEM) || !mip_ptr) {
+               /* out of memory - use dummy tile memory */
+               /* Note if using PERF_TEX_MEM will also skip tile conversion */
+               jit_tex->base = lp_dummy_tile;
+               jit_tex->width = TILE_SIZE/8;
+               jit_tex->height = TILE_SIZE/8;
+               jit_tex->depth = 1;
+               jit_tex->first_level = 0;
+               jit_tex->last_level = 0;
+               jit_tex->mip_offsets[0] = 0;
+               jit_tex->row_stride[0] = 0;
+               jit_tex->img_stride[0] = 0;
+            }
+            else {
+               jit_tex->width = res->width0;
+               jit_tex->height = res->height0;
+               jit_tex->depth = res->depth0;
+               jit_tex->first_level = first_level;
+               jit_tex->last_level = last_level;
+
+               if (llvmpipe_resource_is_texture(res)) {
+                  for (j = first_level; j <= last_level; j++) {
+                     mip_ptr = llvmpipe_get_texture_image_all(lp_tex, j,
+                                                              LP_TEX_USAGE_READ,
+                                                              LP_TEX_LAYOUT_LINEAR);
+                     jit_tex->mip_offsets[j] = (uint8_t *)mip_ptr - (uint8_t *)jit_tex->base;
+                     /*
+                      * could get mip offset directly but need call above to
+                      * invoke tiled->linear conversion.
+                      */
+                     assert(lp_tex->linear_mip_offsets[j] == jit_tex->mip_offsets[j]);
+                     jit_tex->row_stride[j] = lp_tex->row_stride[j];
+                     jit_tex->img_stride[j] = lp_tex->img_stride[j];
+                  }
+
+                 /* FIXME: This is incorrect, as currently layer stride depends on miplevel */
+#if 0
+                  /*
+                   * We don't use anything like first_element (for buffers) or
+                   * first_layer (for arrays), instead adjust the last_element
+                   * (width) or last_layer (depth) plus the base pointer.
+                   * Less parameters and faster at shader execution.
+                   * XXX Could do the same for mip levels.
+                   */
+                  if (res->target == PIPE_TEXTURE_1D_ARRAY ||
+                      res->target == PIPE_TEXTURE_2D_ARRAY) {
+                     jit_tex->depth = view->u.tex.last_layer - view->u.tex.first_layer + 1;
+                     jit_tex->base = (uint8_t *)jit_tex->base +
+                                     view->u.tex.first_layer * lp_tex->img_stride[0];
+                     assert(view->u.tex.first_layer <= view->u.tex.last_layer);
+                     assert(view->u.tex.last_layer < res->array_size);
+                  }
+#endif
+               }
+               else {
+                  unsigned view_blocksize = util_format_get_blocksize(view->format);
+                  /* probably don't really need to fill that out */
+                  jit_tex->mip_offsets[0] = 0;
+                  jit_tex->row_stride[0] = 0;
+                  jit_tex->row_stride[0] = 0;
+
+                  /* everything specified in number of elements here. */
+                  jit_tex->width = view->u.buf.last_element - view->u.buf.first_element + 1;
+                  jit_tex->base = (uint8_t *)jit_tex->base + view->u.buf.first_element *
+                                  view_blocksize;
+                  /* XXX Unsure if we need to sanitize parameters? */
+                  assert(view->u.buf.first_element <= view->u.buf.last_element);
+                  assert(view->u.buf.last_element * view_blocksize < res->width0);
                }
             }
          }
@@ -641,14 +789,18 @@ lp_setup_set_fragment_sampler_views(struct lp_setup_context *setup,
             /*
              * XXX: Where should this be unmapped?
              */
-
-            struct llvmpipe_screen *screen = llvmpipe_screen(tex->screen);
+            struct llvmpipe_screen *screen = llvmpipe_screen(res->screen);
             struct sw_winsys *winsys = screen->winsys;
-            jit_tex->data[0] = winsys->displaytarget_map(winsys, lp_tex->dt,
-                                                        PIPE_TRANSFER_READ);
+            jit_tex->base = winsys->displaytarget_map(winsys, lp_tex->dt,
+                                                         PIPE_TRANSFER_READ);
             jit_tex->row_stride[0] = lp_tex->row_stride[0];
             jit_tex->img_stride[0] = lp_tex->img_stride[0];
-            assert(jit_tex->data[0]);
+            jit_tex->mip_offsets[0] = 0;
+            jit_tex->width = res->width0;
+            jit_tex->height = res->height0;
+            jit_tex->depth = res->depth0;
+            jit_tex->first_level = jit_tex->last_level = 0;
+            assert(jit_tex->base);
          }
       }
    }
@@ -657,6 +809,38 @@ lp_setup_set_fragment_sampler_views(struct lp_setup_context *setup,
 }
 
 
+/**
+ * Called during state validation when LP_NEW_SAMPLER is set.
+ */
+void
+lp_setup_set_fragment_sampler_state(struct lp_setup_context *setup,
+                                    unsigned num,
+                                    struct pipe_sampler_state **samplers)
+{
+   unsigned i;
+
+   LP_DBG(DEBUG_SETUP, "%s\n", __FUNCTION__);
+
+   assert(num <= PIPE_MAX_SAMPLERS);
+
+   for (i = 0; i < PIPE_MAX_SAMPLERS; i++) {
+      const struct pipe_sampler_state *sampler = i < num ? samplers[i] : NULL;
+
+      if (sampler) {
+         struct lp_jit_sampler *jit_sam;
+         jit_sam = &setup->fs.current.jit_context.samplers[i];
+
+         jit_sam->min_lod = sampler->min_lod;
+         jit_sam->max_lod = sampler->max_lod;
+         jit_sam->lod_bias = sampler->lod_bias;
+         COPY_4V(jit_sam->border_color, sampler->border_color.f);
+      }
+   }
+
+   setup->dirty |= LP_SETUP_NEW_FS;
+}
+
+
 /**
  * Is the given texture referenced by any scene?
  * Note: we have to check all scenes including any scenes currently
@@ -671,170 +855,253 @@ lp_setup_is_resource_referenced( const struct lp_setup_context *setup,
    /* check the render targets */
    for (i = 0; i < setup->fb.nr_cbufs; i++) {
       if (setup->fb.cbufs[i]->texture == texture)
-         return PIPE_REFERENCED_FOR_READ | PIPE_REFERENCED_FOR_WRITE;
+         return LP_REFERENCED_FOR_READ | LP_REFERENCED_FOR_WRITE;
    }
    if (setup->fb.zsbuf && setup->fb.zsbuf->texture == texture) {
-      return PIPE_REFERENCED_FOR_READ | PIPE_REFERENCED_FOR_WRITE;
+      return LP_REFERENCED_FOR_READ | LP_REFERENCED_FOR_WRITE;
    }
 
    /* check textures referenced by the scene */
    for (i = 0; i < Elements(setup->scenes); i++) {
       if (lp_scene_is_resource_referenced(setup->scenes[i], texture)) {
-         return PIPE_REFERENCED_FOR_READ;
+         return LP_REFERENCED_FOR_READ;
       }
    }
 
-   return PIPE_UNREFERENCED;
+   return LP_UNREFERENCED;
 }
 
 
 /**
  * Called by vbuf code when we're about to draw something.
  */
-void
-lp_setup_update_state( struct lp_setup_context *setup )
+static boolean
+try_update_scene_state( struct lp_setup_context *setup )
 {
-   struct lp_scene *scene;
-
-   LP_DBG(DEBUG_SETUP, "%s\n", __FUNCTION__);
-
-   setup_check_scene_size_and_flush(setup);
-
-   scene = lp_setup_get_current_scene(setup);
-
-   assert(setup->fs.current.variant);
+   boolean new_scene = (setup->fs.stored == NULL);
+   struct lp_scene *scene = setup->scene;
+   unsigned i;
 
-   /* Some of the 'draw' pipeline stages may have changed some driver state.
-    * Make sure we've processed those state changes before anything else.
-    *
-    * XXX this is the only place where llvmpipe_context is used in the
-    * setup code.  This may get refactored/changed...
-    */
-   {
-      struct llvmpipe_context *lp = llvmpipe_context(scene->pipe);
-      if (lp->dirty) {
-         llvmpipe_update_derived(lp);
-      }
-      assert(lp->dirty == 0);
-   }
+   assert(scene);
 
    if(setup->dirty & LP_SETUP_NEW_BLEND_COLOR) {
       uint8_t *stored;
+      float* fstored;
       unsigned i, j;
+      unsigned size;
 
-      stored = lp_scene_alloc_aligned(scene, 4 * 16, 16);
+      /* Alloc u8_blend_color (16 x i8) and f_blend_color (4 or 8 x f32) */
+      size  = 4 * 16 * sizeof(uint8_t);
+      size += (LP_MAX_VECTOR_LENGTH / 4) * sizeof(float);
+      stored = lp_scene_alloc_aligned(scene, size, LP_MAX_VECTOR_LENGTH);
 
-      if (stored) {
-         /* smear each blend color component across 16 ubyte elements */
-         for (i = 0; i < 4; ++i) {
-            uint8_t c = float_to_ubyte(setup->blend_color.current.color[i]);
-            for (j = 0; j < 16; ++j)
-               stored[i*16 + j] = c;
-         }
+      if (!stored) {
+         assert(!new_scene);
+         return FALSE;
+      }
 
-         setup->blend_color.stored = stored;
+      /* Store floating point colour */
+      fstored = (float*)(stored + 4*16);
+      for (i = 0; i < (LP_MAX_VECTOR_LENGTH / 4); ++i) {
+         fstored[i] = setup->blend_color.current.color[i % 4];
+      }
 
-         setup->fs.current.jit_context.blend_color = setup->blend_color.stored;
+      /* smear each blend color component across 16 ubyte elements */
+      for (i = 0; i < 4; ++i) {
+         uint8_t c = float_to_ubyte(setup->blend_color.current.color[i]);
+         for (j = 0; j < 16; ++j)
+            stored[i*16 + j] = c;
       }
 
+      setup->blend_color.stored = stored;
+      setup->fs.current.jit_context.u8_blend_color = stored;
+      setup->fs.current.jit_context.f_blend_color = fstored;
       setup->dirty |= LP_SETUP_NEW_FS;
    }
 
-   if (setup->dirty & LP_SETUP_NEW_SCISSOR) {
-      float *stored;
-
-      stored = lp_scene_alloc_aligned(scene, 4 * sizeof(int32_t), 16);
-
-      if (stored) {
-         stored[0] = (float) setup->scissor.current.minx;
-         stored[1] = (float) setup->scissor.current.miny;
-         stored[2] = (float) setup->scissor.current.maxx;
-         stored[3] = (float) setup->scissor.current.maxy;
-
-         setup->scissor.stored = stored;
-
-         setup->fs.current.jit_context.scissor_xmin = stored[0];
-         setup->fs.current.jit_context.scissor_ymin = stored[1];
-         setup->fs.current.jit_context.scissor_xmax = stored[2];
-         setup->fs.current.jit_context.scissor_ymax = stored[3];
-      }
+   if (setup->dirty & LP_SETUP_NEW_CONSTANTS) {
+      for (i = 0; i < Elements(setup->constants); ++i) {
+         struct pipe_resource *buffer = setup->constants[i].current.buffer;
+         const unsigned current_size = setup->constants[i].current.buffer_size;
+         const ubyte *current_data = NULL;
 
-      setup->dirty |= LP_SETUP_NEW_FS;
-   }
+         if (buffer) {
+            /* resource buffer */
+            current_data = (ubyte *) llvmpipe_resource_data(buffer);
+         }
+         else if (setup->constants[i].current.user_buffer) {
+            /* user-space buffer */
+            current_data = (ubyte *) setup->constants[i].current.user_buffer;
+         }
 
-   if(setup->dirty & LP_SETUP_NEW_CONSTANTS) {
-      struct pipe_resource *buffer = setup->constants.current;
+         if (current_data) {
+            current_data += setup->constants[i].current.buffer_offset;
 
-      if(buffer) {
-         unsigned current_size = buffer->width0;
-         const void *current_data = llvmpipe_resource_data(buffer);
+            /* TODO: copy only the actually used constants? */
 
-         /* TODO: copy only the actually used constants? */
+            if (setup->constants[i].stored_size != current_size ||
+               !setup->constants[i].stored_data ||
+               memcmp(setup->constants[i].stored_data,
+                      current_data,
+                      current_size) != 0) {
+               void *stored;
 
-         if(setup->constants.stored_size != current_size ||
-            !setup->constants.stored_data ||
-            memcmp(setup->constants.stored_data,
-                   current_data,
-                   current_size) != 0) {
-            void *stored;
+               stored = lp_scene_alloc(scene, current_size);
+               if (!stored) {
+                  assert(!new_scene);
+                  return FALSE;
+               }
 
-            stored = lp_scene_alloc(scene, current_size);
-            if(stored) {
                memcpy(stored,
                       current_data,
                       current_size);
-               setup->constants.stored_size = current_size;
-               setup->constants.stored_data = stored;
+               setup->constants[i].stored_size = current_size;
+               setup->constants[i].stored_data = stored;
             }
          }
-      }
-      else {
-         setup->constants.stored_size = 0;
-         setup->constants.stored_data = NULL;
-      }
+         else {
+            setup->constants[i].stored_size = 0;
+            setup->constants[i].stored_data = NULL;
+         }
 
-      setup->fs.current.jit_context.constants = setup->constants.stored_data;
-      setup->dirty |= LP_SETUP_NEW_FS;
+         setup->fs.current.jit_context.constants[i] = setup->constants[i].stored_data;
+         setup->dirty |= LP_SETUP_NEW_FS;
+      }
    }
 
 
-   if(setup->dirty & LP_SETUP_NEW_FS) {
-      if(!setup->fs.stored ||
-         memcmp(setup->fs.stored,
-                &setup->fs.current,
-                sizeof setup->fs.current) != 0) {
+   if (setup->dirty & LP_SETUP_NEW_FS) {
+      if (!setup->fs.stored ||
+          memcmp(setup->fs.stored,
+                 &setup->fs.current,
+                 sizeof setup->fs.current) != 0)
+      {
+         struct lp_rast_state *stored;
+         
          /* The fs state that's been stored in the scene is different from
           * the new, current state.  So allocate a new lp_rast_state object
           * and append it to the bin's setup data buffer.
           */
-         uint i;
-         struct lp_rast_state *stored =
-            (struct lp_rast_state *) lp_scene_alloc(scene, sizeof *stored);
-         if(stored) {
-            memcpy(stored,
-                   &setup->fs.current,
-                   sizeof setup->fs.current);
-            setup->fs.stored = stored;
-
-            /* put the state-set command into all bins */
-            lp_scene_bin_state_command( scene,
-                                       lp_rast_set_state, 
-                                       lp_rast_arg_state(setup->fs.stored) );
+         stored = (struct lp_rast_state *) lp_scene_alloc(scene, sizeof *stored);
+         if (!stored) {
+            assert(!new_scene);
+            return FALSE;
          }
 
+         memcpy(stored,
+                &setup->fs.current,
+                sizeof setup->fs.current);
+         setup->fs.stored = stored;
+         
          /* The scene now references the textures in the rasterization
           * state record.  Note that now.
           */
          for (i = 0; i < Elements(setup->fs.current_tex); i++) {
-            if (setup->fs.current_tex[i])
-               lp_scene_add_resource_reference(scene, setup->fs.current_tex[i]);
+            if (setup->fs.current_tex[i]) {
+               if (!lp_scene_add_resource_reference(scene,
+                                                    setup->fs.current_tex[i],
+                                                    new_scene)) {
+                  assert(!new_scene);
+                  return FALSE;
+               }
+            }
          }
       }
    }
 
+   if (setup->dirty & LP_SETUP_NEW_SCISSOR) {
+      setup->draw_region = setup->framebuffer;
+      if (setup->scissor_test) {
+         u_rect_possible_intersection(&setup->scissor,
+                                      &setup->draw_region);
+      }
+      /* If the framebuffer is large we have to think about fixed-point
+       * integer overflow.  For 2K by 2K images, coordinates need 15 bits
+       * (2^11 + 4 subpixel bits).  The product of two such numbers would
+       * use 30 bits.  Any larger and we could overflow a 32-bit int.
+       *
+       * To cope with this problem we check if triangles are large and
+       * subdivide them if needed.
+       */
+      setup->subdivide_large_triangles = (setup->fb.width > 2048 &&
+                                          setup->fb.height > 2048);
+   }
+                                      
    setup->dirty = 0;
 
    assert(setup->fs.stored);
+   return TRUE;
+}
+
+boolean
+lp_setup_update_state( struct lp_setup_context *setup,
+                       boolean update_scene )
+{
+   /* Some of the 'draw' pipeline stages may have changed some driver state.
+    * Make sure we've processed those state changes before anything else.
+    *
+    * XXX this is the only place where llvmpipe_context is used in the
+    * setup code.  This may get refactored/changed...
+    */
+   {
+      struct llvmpipe_context *lp = llvmpipe_context(setup->pipe);
+      if (lp->dirty) {
+         llvmpipe_update_derived(lp);
+      }
+
+      if (lp->setup->dirty) {
+         llvmpipe_update_setup(lp);
+      }
+
+      assert(setup->setup.variant);
+
+      /* Will probably need to move this somewhere else, just need  
+       * to know about vertex shader point size attribute.
+       */
+      setup->psize = lp->psize_slot;
+
+      assert(lp->dirty == 0);
+
+      assert(lp->setup_variant.key.size == 
+            setup->setup.variant->key.size);
+
+      assert(memcmp(&lp->setup_variant.key,
+                   &setup->setup.variant->key,
+                   setup->setup.variant->key.size) == 0);
+   }
+
+   if (update_scene && setup->state != SETUP_ACTIVE) {
+      if (!set_scene_state( setup, SETUP_ACTIVE, __FUNCTION__ ))
+         return FALSE;
+   }
+
+   /* Only call into update_scene_state() if we already have a
+    * scene:
+    */
+   if (update_scene && setup->scene) {
+      assert(setup->state == SETUP_ACTIVE);
+
+      if (try_update_scene_state(setup))
+         return TRUE;
+
+      /* Update failed, try to restart the scene.
+       *
+       * Cannot call lp_setup_flush_and_restart() directly here
+       * because of potential recursion.
+       */
+      if (!set_scene_state(setup, SETUP_FLUSHED, __FUNCTION__))
+         return FALSE;
+
+      if (!set_scene_state(setup, SETUP_ACTIVE, __FUNCTION__))
+         return FALSE;
+
+      if (!setup->scene)
+         return FALSE;
+
+      return try_update_scene_state(setup);
+   }
+
+   return TRUE;
 }
 
 
@@ -846,7 +1113,7 @@ lp_setup_destroy( struct lp_setup_context *setup )
 {
    uint i;
 
-   reset_context( setup );
+   lp_setup_reset( setup );
 
    util_unreference_framebuffer_state(&setup->fb);
 
@@ -854,19 +1121,21 @@ lp_setup_destroy( struct lp_setup_context *setup )
       pipe_resource_reference(&setup->fs.current_tex[i], NULL);
    }
 
-   pipe_resource_reference(&setup->constants.current, NULL);
+   for (i = 0; i < Elements(setup->constants); i++) {
+      pipe_resource_reference(&setup->constants[i].current.buffer, NULL);
+   }
 
    /* free the scenes in the 'empty' queue */
-   while (1) {
-      struct lp_scene *scene = lp_scene_dequeue(setup->empty_scenes, FALSE);
-      if (!scene)
-         break;
+   for (i = 0; i < Elements(setup->scenes); i++) {
+      struct lp_scene *scene = setup->scenes[i];
+
+      if (scene->fence)
+         lp_fence_wait(scene->fence);
+
       lp_scene_destroy(scene);
    }
 
-   lp_scene_queue_destroy(setup->empty_scenes);
-
-   lp_rast_destroy( setup->rast );
+   lp_fence_reference(&setup->last_fence, NULL);
 
    FREE( setup );
 }
@@ -882,37 +1151,36 @@ lp_setup_create( struct pipe_context *pipe,
                  struct draw_context *draw )
 {
    struct llvmpipe_screen *screen = llvmpipe_screen(pipe->screen);
-   struct lp_setup_context *setup = CALLOC_STRUCT(lp_setup_context);
+   struct lp_setup_context *setup;
    unsigned i;
 
-   if (!setup)
-      return NULL;
+   setup = CALLOC_STRUCT(lp_setup_context);
+   if (!setup) {
+      goto no_setup;
+   }
 
    lp_setup_init_vbuf(setup);
+   
+   /* Used only in update_state():
+    */
+   setup->pipe = pipe;
 
-   setup->empty_scenes = lp_scene_queue_create();
-   if (!setup->empty_scenes)
-      goto fail;
 
-   /* XXX: move this to the screen and share between contexts:
-    */
    setup->num_threads = screen->num_threads;
-   setup->rast = lp_rast_create(screen->num_threads);
-   if (!setup->rast) 
-      goto fail;
-
    setup->vbuf = draw_vbuf_stage(draw, &setup->base);
-   if (!setup->vbuf)
-      goto fail;
+   if (!setup->vbuf) {
+      goto no_vbuf;
+   }
 
    draw_set_rasterize_stage(draw, setup->vbuf);
    draw_set_render(draw, &setup->base);
 
    /* create some empty scenes */
    for (i = 0; i < MAX_SCENES; i++) {
-      setup->scenes[i] = lp_scene_create( pipe, setup->empty_scenes );
-
-      lp_scene_enqueue(setup->empty_scenes, setup->scenes[i]);
+      setup->scenes[i] = lp_scene_create( pipe );
+      if (!setup->scenes[i]) {
+         goto no_scenes;
+      }
    }
 
    setup->triangle = first_triangle;
@@ -923,17 +1191,17 @@ lp_setup_create( struct pipe_context *pipe,
 
    return setup;
 
-fail:
-   if (setup->rast)
-      lp_rast_destroy( setup->rast );
-   
-   if (setup->vbuf)
-      ;
-
-   if (setup->empty_scenes)
-      lp_scene_queue_destroy(setup->empty_scenes);
+no_scenes:
+   for (i = 0; i < MAX_SCENES; i++) {
+      if (setup->scenes[i]) {
+         lp_scene_destroy(setup->scenes[i]);
+      }
+   }
 
+   setup->vbuf->destroy(setup->vbuf);
+no_vbuf:
    FREE(setup);
+no_setup:
    return NULL;
 }
 
@@ -945,22 +1213,33 @@ void
 lp_setup_begin_query(struct lp_setup_context *setup,
                      struct llvmpipe_query *pq)
 {
-   struct lp_scene * scene = lp_setup_get_current_scene(setup);
-   union lp_rast_cmd_arg cmd_arg;
-
    /* init the query to its beginning state */
-   pq->done = FALSE;
-   pq->tile_count = 0;
-   pq->num_tiles = scene->tiles_x * scene->tiles_y;
-   assert(pq->num_tiles > 0);
+   assert(setup->active_query[pq->type] == NULL);
 
-   memset(pq->count, 0, sizeof(pq->count));  /* reset all counters */
+   set_scene_state(setup, SETUP_ACTIVE, "begin_query");
+   
+   setup->active_query[pq->type] = pq;
+
+   /* XXX: It is possible that a query is created before the scene
+    * has been created. This means that setup->scene == NULL resulting
+    * in the query not being binned and thus is ignored.
+    */
+
+   if (setup->scene) {
+      if (!lp_scene_bin_everywhere(setup->scene,
+                                   LP_RAST_OP_BEGIN_QUERY,
+                                   lp_rast_arg_query(pq))) {
 
-   set_scene_state( setup, SETUP_ACTIVE );
+         if (!lp_setup_flush_and_restart(setup))
+            return;
 
-   cmd_arg.query_obj = pq;
-   lp_scene_bin_everywhere(scene, lp_rast_begin_query, cmd_arg);
-   pq->binned = TRUE;
+         if (!lp_scene_bin_everywhere(setup->scene,
+                                      LP_RAST_OP_BEGIN_QUERY,
+                                      lp_rast_arg_query(pq))) {
+            return;
+         }
+      }
+   }
 }
 
 
@@ -970,11 +1249,50 @@ lp_setup_begin_query(struct lp_setup_context *setup,
 void
 lp_setup_end_query(struct lp_setup_context *setup, struct llvmpipe_query *pq)
 {
-   struct lp_scene * scene = lp_setup_get_current_scene(setup);
-   union lp_rast_cmd_arg cmd_arg;
+   set_scene_state(setup, SETUP_ACTIVE, "end_query");
+
+   if (pq->type != PIPE_QUERY_TIMESTAMP) {
+      assert(setup->active_query[pq->type] == pq);
+      setup->active_query[pq->type] = NULL;
+   }
+
+   /* Setup will automatically re-issue any query which carried over a
+    * scene boundary, and the rasterizer automatically "ends" queries
+    * which are active at the end of a scene, so there is no need to
+    * retry this commands on failure.
+    */
+   if (setup->scene) {
+      /* pq->fence should be the fence of the *last* scene which
+       * contributed to the query result.
+       */
+      lp_fence_reference(&pq->fence, setup->scene->fence);
+
+      if (!lp_scene_bin_everywhere(setup->scene,
+                                   LP_RAST_OP_END_QUERY,
+                                   lp_rast_arg_query(pq))) {
+         lp_setup_flush(setup, NULL, __FUNCTION__);
+      }
+   }
+   else {
+      lp_fence_reference(&pq->fence, setup->last_fence);
+   }
+}
+
+
+boolean
+lp_setup_flush_and_restart(struct lp_setup_context *setup)
+{
+   if (0) debug_printf("%s\n", __FUNCTION__);
 
-   set_scene_state( setup, SETUP_ACTIVE );
+   assert(setup->state == SETUP_ACTIVE);
 
-   cmd_arg.query_obj = pq;
-   lp_scene_bin_everywhere(scene, lp_rast_end_query, cmd_arg);
+   if (!set_scene_state(setup, SETUP_FLUSHED, __FUNCTION__))
+      return FALSE;
+   
+   if (!lp_setup_update_state(setup, TRUE))
+      return FALSE;
+
+   return TRUE;
 }
+
+